<p>The aluminum industry, characterized by extreme operational environments involving high temperatures and intensive energy consumption, faces significant challenges in real-time optimization owing to complex multi-variable coupling and nonlinear dynamics. This research delineates the evolutionary trajectory of Artificial Intelligence (AI) applications across the entire value chain, encompassing alumina production, aluminum electrolysis, and materials processing. By analyzing the transition from single-parameter heuristics to sophisticated mechanism-data fusion models, this study elucidates how the integration of physicochemical principles into data-driven frameworks overcomes the inherent limitations of traditional mechanistic models. Specifically, the implementation of spatio-temporal feature mining and multi-modal fusion has revolutionized anomaly detection and quality prediction in industrial settings. Despite these advancements, the study identifies critical bottlenecks in data integrity, model interpretability, and cross-domain transferability. To bridge these gaps, a comprehensive technical roadmap is proposed, integrating interpretable AI architectures and industrial large-scale models, thereby providing a systemic foundation for the digital and intelligent metamorphosis of the global aluminum sector.</p> Graphical Abstract <p></p>